deniz.in

Markets

Weather

Loading weather

· via Hacker News – Front Page (hnrss.org)

Pulsar MCP server lets AI agents in different coding harnesses message each other

Pulsar is an open-source MCP server that lets AI coding agents running in Claude Code, Codex, Cursor, Gemini CLI and other harnesses discover, address and message each other on one machine, with push delivery that wakes idle agents.

Pulsar MCP server lets AI agents in different coding harnesses message each other

A new open-source project called Pulsar, featured on Hacker News on 17 November, lets AI coding agents discover and message each other even when they run in different harnesses. According to the project's README by developer yocontra, Pulsar is an MCP server that places every coding agent on a machine — Claude Code, Codex, OpenCode, Cursor, Gemini CLI, pi and omp — onto a shared message bus, so a Claude session can reach a Codex session or a Cursor agent as easily as it calls a tool.

One registry per machine

A small daemon runs per machine, started on demand and stopped when idle. It keeps the agent registry and writes a SQLite message log under ~/.local/state/pulsar, and each agent's MCP server is a thin client speaking to it over a Unix socket. The daemon also scans for live Claude Code and OpenCode sessions that never loaded Pulsar, so those sessions still appear in listings and can receive messages. Claude sessions keep the names Claude assigns; other agents appear as harness@project.

Agents are addressable by role, project or PR

Agents automatically attach metadata: harness, model, project, git branch, worktree and the pull request they work on, which the README says is derived offline from local git refs and session logs. They can add their own tags, such as role:reviewer or issue:1378, plus a free-text activity line. Messages can be addressed to a name, a selector like project=apparish or pr=1402 role=reviewer, a text search, or * to reach every agent on the machine. The README's example of messaging your reviewer therefore reduces to a single lookup.

Push delivery and blocking primitives

Delivery is push-based rather than polled: a normal-priority message wakes an idle agent, and urgent messages can land mid-turn where the harness supports it. Each harness gets a dedicated push path — Claude Code's per-session inbox socket, Codex's queue --resume mechanism, OpenCode's local prompt endpoint, and a companion extension for pi and omp. Generic MCP clients without a push path fall back to an inbox, with unread counts attached to every tool result. On top of that sit reply (threaded), ask (blocks until an answer arrives) and wait (blocks until a message, a reply, or another agent going idle), which the project positions as an alternative to busy-waiting. A CLI mirrors the tools with commands like pulsar ls, pulsar send and pulsar tail.

Safety boundaries

The socket lives in a 0700 directory, making the user account the trust boundary, as it already is for ordinary files. The daemon stamps each message with the sender's connection, so an agent cannot claim another identity, and no agent may use the name "operator". Incoming text is labelled as coming from a peer rather than the human, cannot grant permissions, and each harness's own guardrails still apply on top. Agents can accept, hold or refuse inbound messages, and the project reports rate limits and no telemetry.

Installation and current limits

Running npx pulsar@latest setup detects installed harnesses, shows exactly which files it will change, asks once, and backs up anything it touches. A curl script installs a private Node runtime for machines without Node, and harness-native plugin commands exist for Claude Code, Codex, pi and omp; any MCP client can point at npx -y pulsar mcp. Version 1 is single-machine. The README contrasts it with Claude Code's built-in SendMessage, which only works between Claude sessions, and lists cross-machine messaging plus remote control over Tailscale as upcoming features.

Why it matters

Multi-agent coding workflows have largely been single-vendor affairs: Claude Code's own messaging, for instance, stops at Claude agents. Pulsar treats the harness as interchangeable and models agents as addressable peers with roles, projects and availability — closer to how services coordinate in distributed systems than how chat assistants behave. It is also an early, concrete use of MCP for agent-to-agent coordination rather than plain tool access. The conservative defaults — peer-labelled, non-authorizing, rate-limited messages — acknowledge that inter-agent text is an injection surface, and that discipline will matter as fleets of concurrently running agents become routine.

  • #mcp
  • #ai-agents
  • #developer-tools
  • #open-source
  • #multi-agent

Related posts